Built-in Certificates and Automated Verification of Learning-Based Control
Abstract
The goal of the proposed research is to develop built-in robustness certificates and automated verification procedures for learning-based control of uncertain nonlinear systems. Recently, machine learning techniques have shown great promise for advanced control tasks involving complex nonlinear dynamics (e.g., turbulence effects, contact forces, etc) and-or high-dimensional sensory measurements (e.g., images). On one hand, the combination of neural net-works and model-free reinforcement learning has led to remarkable performance gains in various demos and showcases. On the other hand, it remains unclear how to get such learning-based controllers properly certified such that they can be safely deployed for practical autonomous systems. Analysis techniques currently being used for commercial-military aircraft are not suited to certify learning-based flight control systems with complex neural networks and-or sensory data in the loop. Modern neural network controllers and rich sensing modalities can exhibit highly nonlinear behaviors, causing difficulty for generating tight , or even moderately non-conservative, certificates of large-scale learning-based feed back systems. Our proposed research aims at bridging this gap via the developments of built-in robustness certificates and automated verification methods for learning-based control with neural networks and-or sensory data in the loop. Specifically, we will take an interdisciplinary perspective on nonlinear control, machine learning, and robustness certification, generalizing and integrating techniques including nonlinear stability theorems, mixed integer programming, and robust deep learning.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Mar 14, 2024
- Source ID
- FA95502310732
Entities
People
- Bin Hu
Organizations
- Air Force Office of Scientific Research
- United States Air Force
- University of Illinois Urbana–Champaign